im honestly so sick of the "skill issue just prompt better" copium whenever an ai agent starts churning out pure slop after like 20 turns. tbh i finally audited my api logs this week bc my anthropic bill was exploding for no reason and realized something that actually pissed me off.

the models arent actually losing their minds. they are literally just suffocating on their own context window before they even attempt to reason or write code.

if u watch what these agents actually do on any repo over 10k lines its insane

blind exploration. they just recursively grep and read like 40 files to find one function. half the time instead of finding my existing ui component it just hallucinates a completely duplicate one from scratch lmao

raw ingestion. itll read a massive 2k line file just to update a 5 line interface... why

shell & tool diarrhea. verbose test logs and bloated mcp tool definitions are eating like 30k tokens before the agent even types a single line

absolute goldfish memory. every session is groundhog day. it just re-reads the same exact files bc it has zero project aware memory

once the context window gets to like 80% full of this pure noise the agents iq visibly drops to room temp and the architectural decay starts. standard rag or compressing outputs doesnt fix this at all. the agent is fundamentally blind to how a codebase is actually structured until it burns through your wallet reading raw text.

are we all really just accepting this weird productivity paradox where we save an hour of typing just to spend 5 hours fixing the architectural spaghetti the ai just made?? do we need some ground up new agent that actually understands code as a graph before wasting tokens reading raw text? or am i literally the only one dealing with this

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.64